Chao Xing 2015-09-28
From cslt Wiki
Last Week Solution: 1. DNN program has some problems, fix these & DNN contain same performance to Linear. Problem is training process, training samples are equal to speakers' so this will lead model less compute 50% training samples. Plan to do: 1. Help Lantian finish his papers. 2. Run for a small vocabulary size, such as 5000. Problem: 1. DNN still not better than linear model.
Solution: 1. DNN with dropout method, and add SAA to reduce randomly properties. Plan to do: 1. Run for a small vocabulary size, such as 5000. 2. Run and test some ideas.